Abstract
Brain imaging genetics becomes an important research topic since it can reveal complex associations between genetic factors and the structures or functions of the human brain. Sparse canonical correlation analysis (SCCA) is a popular bi-multivariate association identification method. To mine the complex genetic basis of brain imaging phenotypes, there arise many SCCA methods with a variety of norms for incorporating different structures of interest. They often use the group lasso penalty, the fused lasso or the graph/network guided fused lasso ones. However, the group lasso methods have limited capability because of the incomplete or unavailable prior knowledge in real applications. The fused lasso and graph/network guided methods are sensitive to the sign of the sample correlation which may be incorrectly estimated. In this paper, we introduce two new penalties to improve the fused lasso and the graph/network guided lasso penalties in structured sparse learning. We impose both penalties to the SCCA model and propose an optimization algorithm to solve it. The proposed SCCA method has a strong upper bound of grouping effects for both positively and negatively highly correlated variables. We show that, on both synthetic and real neuroimaging genetics data, the proposed SCCA method performs better than or equally to the conventional methods using fused lasso or graph/network guided fused lasso. In particular, the proposed method identifies higher canonical correlation coefficients and captures clearer canonical weight patterns, demonstrating its promising capability in revealing biologically meaningful imaging genetic associations.